pith. sign in

arxiv: 2605.01856 · v1 · submitted 2026-05-03 · 📊 stat.ML · cs.LG· stat.ME

Stable Blanket with Hidden Variables and Cycles

Hanqing Xiang This is my paper

Pith reviewed 2026-05-09 16:52 UTC · model grok-4.3

classification 📊 stat.ML cs.LGstat.ME
keywords stable blanketMarkov blankethidden variablescausal cyclesADMGsDMGsstabilized regressiongraphical causal models
0
0 comments X

The pith

Graphical criteria identify stable predictor sets even when models include hidden variables and causal cycles.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that Markov blankets and stable blankets for regression can be characterized using mixed graphs in settings with latent variables, feedback loops, or both. This extends prior work limited to acyclic fully observed systems, allowing identification of predictor sets whose conditional link to a response remains unchanged under interventions. The results specify when such sets are minimal or unique by tracking how interventions propagate through sub-districts or strongly connected components. A reader would care because real data routinely features unmeasured factors and reciprocal effects that alter which variables qualify as stable.

Core claim

In acyclic directed mixed graphs, m-separation and intervened sub-districts characterize Markov blankets and stable frontiers. In directed mixed graphs with cycles, σ-separation treats strongly connected components as units to find stable blankets. Combining both handles models with hidden variables and cycles simultaneously, yielding conditions under which the response is conditionally independent of intervention variables given a suitable predictor set.

What carries the argument

Intervened sub-districts in ADMGs and strongly connected components in DMGs, which track intervention propagation and enable separation-based identification of stable blankets.

Load-bearing premise

That m-separation in ADMGs and σ-separation in DMGs correctly represent the conditional independencies created by hidden variables and cycles.

What would settle it

A concrete graph containing a hidden variable or cycle in which a set identified by the criteria changes its conditional relationship to the response after an intervention that the theory predicts should leave it invariant.

Figures

Figures reproduced from arXiv: 2605.01856 by Hanqing Xiang.

Figure 1
Figure 1. Figure 1: A DAG without hidden variables. The intervention nodes are view at source ↗
Figure 2
Figure 2. Figure 2: Two motivating examples. Left: a model with a hidden variable view at source ↗
Figure 3
Figure 3. Figure 3: ADMG generated by latent projection with respect to hidden variables. view at source ↗
Figure 4
Figure 4. Figure 4: Graphical illustration for condition (2) in Theorem view at source ↗
Figure 5
Figure 5. Figure 5: Graphical model with cycles but without hidden variables. view at source ↗
Figure 6
Figure 6. Figure 6: Graphical illustration for the relative, where i is a relative of j. view at source ↗
Figure 7
Figure 7. Figure 7: Cyclic DMG generated by latent projection with respect to hidden variables. view at source ↗
read the original abstract

Stabilized regression aims to identify a set of predictors whose conditional relationship with a response variable remains invariant across different environments. Existing graphical characterizations of the stable blanket are mainly developed for structural causal models (SCMs) without hidden variables or causal cycles. However, latent variables and feedback relationships naturally arise in many applications, and they can change both the Markov blanket and the set of predictors that remain stable under interventions. This paper studies stable blankets in graphical causal models with hidden variables, causal cycles, and both features simultaneously. For models with hidden variables, we use acyclic directed mixed graphs (ADMGs) and $m$-separation to characterize the Markov blanket and to construct intervention-stable predictor sets. We introduce the notion of an intervened sub-district and use it to describe how interventions may affect districts connected to the response. For models with cycles, we work with directed graphs (DGs) and directed mixed graphs (DMGs) together with $\sigma$-separation, treating strongly connected components (SCCs) as the basic graphical units. We then combine these ideas to analyze models with both hidden variables and cycles. The main results give graphical characterizations of Markov blankets, stable frontiers, and stable blankets in these generalized settings. In particular, we identify conditions under which the response is conditionally independent of intervention variables given a suitable predictor set, and we describe when such sets are minimal or unique. These results extend the graphical interpretation of stabilized regression beyond acyclic fully observed models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims to extend graphical characterizations of Markov blankets, stable frontiers, and stable blankets for stabilized regression to structural causal models with hidden variables (using ADMGs and m-separation), causal cycles (using DGs/DMGs and σ-separation), and both features simultaneously. It introduces the intervened sub-district to capture how interventions propagate to districts connected to the response variable and treats strongly connected components (SCCs) as basic units in cyclic models. The main results identify conditions under which the response is conditionally independent of intervention variables given a suitable predictor set and describe when such sets are minimal or unique.

Significance. If the characterizations hold, this is a significant extension of stabilized regression beyond acyclic, fully observed models, as hidden variables and feedback loops are common in applications. The work builds directly on established separation criteria (m-separation and σ-separation) and introduces targeted graphical constructs (intervened sub-districts and SCC units) to handle interventions, which is a strength. It provides concrete conditions for conditional independence and minimality/uniqueness, potentially enabling more reliable predictor selection in complex causal systems.

major comments (2)
  1. [Section on models with both hidden variables and cycles] The central results on stable blankets in the combined hidden-variables-and-cycles setting rely on the claim that m-separation in ADMGs and σ-separation in DMGs, together with the intervened sub-district and SCC-based units, correctly encode the intervention-stable conditional independencies. The manuscript does not supply an explicit verification, proof sketch, or counter-example check that the combined graphical criterion matches the d-separation semantics of the underlying SCM when both latents and cycles are present simultaneously; this is load-bearing for the main results.
  2. [Section introducing intervened sub-district for ADMGs] The definition and properties of the intervened sub-district (introduced to describe intervention effects on districts connected to the response) are used to construct stable predictor sets, but the manuscript does not demonstrate that this construction captures all relevant paths without introducing spurious independencies or missing latent-induced paths in the presence of cycles; this underpins the claimed graphical characterization of stable frontiers.
minor comments (2)
  1. [Abstract] The abstract states that main results exist and lists the graphical tools but supplies no proof outline or illustrative example; adding a one-sentence indication of the key technical step (e.g., how the intervened sub-district is formally defined) would improve readability.
  2. [Preliminaries] Notation for the various graph classes (ADMGs, DMGs, DGs) and separation criteria should be introduced with a brief comparison table or paragraph to avoid confusion when the combined setting is discussed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading of the manuscript and for the positive assessment of its potential significance. We address the two major comments point by point below. Where the comments identify opportunities to strengthen the presentation of the combined hidden-variables-and-cycles results, we will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Section on models with both hidden variables and cycles] The central results on stable blankets in the combined hidden-variables-and-cycles setting rely on the claim that m-separation in ADMGs and σ-separation in DMGs, together with the intervened sub-district and SCC-based units, correctly encode the intervention-stable conditional independencies. The manuscript does not supply an explicit verification, proof sketch, or counter-example check that the combined graphical criterion matches the d-separation semantics of the underlying SCM when both latents and cycles are present simultaneously; this is load-bearing for the main results.

    Authors: We agree that an explicit verification for the joint setting would improve clarity. The characterizations are obtained by composing the established soundness and completeness of m-separation (for ADMGs) and σ-separation (for DMGs) with the intervened-sub-district construction and the treatment of SCCs as atomic units. We will add a concise proof sketch in the appendix that confirms the combined criterion preserves the required conditional independencies, by showing that no new m-separated or σ-separated paths are created or destroyed at the interface between latent-induced edges and cyclic components. revision: yes

  2. Referee: [Section introducing intervened sub-district for ADMGs] The definition and properties of the intervened sub-district (introduced to describe intervention effects on districts connected to the response) are used to construct stable predictor sets, but the manuscript does not demonstrate that this construction captures all relevant paths without introducing spurious independencies or missing latent-induced paths in the presence of cycles; this underpins the claimed graphical characterization of stable frontiers.

    Authors: The intervened sub-district is defined on the mixed graph to isolate intervention effects on districts adjacent to the response while respecting m-separation. When cycles are present, SCCs are treated as single units under σ-separation. We will augment the manuscript with a short demonstration (including a small illustrative example) that the construction enumerates all relevant paths, including those induced by latent variables, and does not introduce spurious independencies; the argument relies on the fact that σ-separation within an SCC is closed under the district-level intervention operation. revision: yes

Circularity Check

0 steps flagged

No significant circularity in graphical characterizations

full rationale

The paper extends standard m-separation on ADMGs and σ-separation on DMGs to define Markov blankets, stable frontiers, and stable blankets under hidden variables and cycles. It introduces intervened sub-districts and SCC-based units as new graphical constructs to track intervention effects. These steps are definitional extensions of existing separation semantics rather than self-referential reductions, fitted-parameter predictions, or load-bearing self-citations. The central claims derive from applying the separation criteria to the augmented graphs, with no equations or results shown to collapse back to their own inputs by construction. The work remains self-contained against external graphical-model benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The paper rests on standard properties of m-separation and σ-separation plus two new graphical constructs (intervened sub-district and SCC units) whose correctness is asserted rather than derived from more primitive axioms.

axioms (2)
  • domain assumption m-separation in ADMGs identifies all conditional independencies in the presence of hidden variables
    Invoked to characterize the Markov blanket and stable predictor sets
  • domain assumption σ-separation in DMGs identifies conditional independencies in the presence of cycles
    Used when treating SCCs as basic units for models with feedback
invented entities (1)
  • intervened sub-district no independent evidence
    purpose: To describe how interventions affect districts connected to the response variable
    New notion introduced to handle intervention effects in the hidden-variable case

pith-pipeline@v0.9.0 · 5554 in / 1513 out tokens · 33507 ms · 2026-05-09T16:52:32.646048+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Prediction-Intervention Games and Invariant Sets

    stat.ML 2026-05 unverdicted novelty 7.0

    In prediction-intervention games, stable-blanket predictors are at least as good as causal-parent predictors for two classes of follower objectives and can be worst-case optimal under additional conditions.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Markov properties for acyclic directed mixed graphs

    Richardson, T., 2003. Markov properties for acyclic directed mixed graphs. Scandinavian Jour- nal of Statistics, 30(1), pp.145-157

    work page 2003

  2. [2]

    and Scholkopf, B., 2017

    Peters, J., Janzing, D. and Scholkopf, B., 2017. Elements of causal inference: foundations and learning algorithms. MIT press

    work page 2017

  3. [3]

    and Bühlmann, P., 2021

    Pfister, N., Williams, E.G., Peters, J., Aebersold, R. and Bühlmann, P., 2021. Stabilizing variable selection and regression. The Annals of Applied Statistics, 15(3), pp.1220-1246

    work page 2021

  4. [4]

    and Spirtes, P., 2002

    Richardson, T. and Spirtes, P., 2002. Ancestral graph Markov models. The Annals of Statistics, 30(4), pp.962-1030

    work page 2002

  5. [5]

    and Cooper, G.F., 2021, December

    Triantafillou, S., Jabbari, F. and Cooper, G.F., 2021, December. Causal and interventional markov boundaries. In Uncertainty in Artificial Intelligence (pp. 1434-1443). PMLR

    work page 2021

  6. [6]

    Causality

    Pearl, J., 2009. Causality. Cambridge university press

    work page 2009

  7. [7]

    and Scheines, R., 2000

    Spirtes, P., Glymour, C.N. and Scheines, R., 2000. Causation, prediction, and search. MIT press. 39

    work page 2000

  8. [8]

    and Mooij, J.M., 2021

    Bongers, S., Forré, P., Peters, J. and Mooij, J.M., 2021. Foundations of structural causal models with cycles and latent variables. The Annals of Statistics, 49(5), pp.2885-2915

    work page 2021

  9. [9]

    Wu, C., Zhao, H., Fang, H.andDeng, M., 2017.Graphicalmodelselection withlatentvariables

    work page 2017

  10. [10]

    Graphical models (Vol

    Lauritzen, S.L., 1996. Graphical models (Vol. 17). Clarendon press

    work page 1996

  11. [11]

    and Willsky, A.S., 2010, September

    Chandrasekaran, V., Parrilo, P.A. and Willsky, A.S., 2010, September. Latent variable graph- ical model selection via convex optimization. In 2010 48th Annual Allerton Conference on Communication, Control, and Computing (Allerton) (pp. 1610-1613). IEEE

    work page 2010

  12. [12]

    and Koutsoukos, X.D., 2010

    Aliferis, C.F., Statnikov, A., Tsamardinos, I., Mani, S. and Koutsoukos, X.D., 2010. Local causal and Markov blanket induction for causal discovery and feature selection for classification part I: algorithms and empirical evaluation. Journal of Machine Learning Research, 11(1)

    work page 2010

  13. [13]

    and Friedman, N., 2009

    Koller, D. and Friedman, N., 2009. Probabilistic graphical models: principles and techniques. MIT press

    work page 2009

  14. [14]

    and Baltieri, M., 2022

    Bruineberg, J., Dołęga, K., Dewhurst, J. and Baltieri, M., 2022. The emperor’s new Markov blankets. Behavioral and Brain Sciences, 45, p.e183

    work page 2022

  15. [15]

    and Elisseeff, A., 2008

    Pellet, J.P. and Elisseeff, A., 2008. Finding latent causes in causal networks: an efficient ap- proach based on Markov blankets. Advances in Neural Information Processing Systems, 21

    work page 2008

  16. [16]

    and Manderick, B., 2007, October

    Meganck, S., Leray, P. and Manderick, B., 2007, October. Causal graphical models with la- tent variables: Learning and inference. In European Conference on Symbolic and Quantitative Approaches to Reasoning and Uncertainty (pp. 5-16). Berlin, Heidelberg: Springer Berlin Hei- delberg

    work page 2007

  17. [17]

    and Peters, J., 2022, August

    Pfister, N. and Peters, J., 2022, August. Identifiability of sparse causal effects using instrumen- tal variables. In Uncertainty in Artificial Intelligence (pp. 1613-1622). PMLR

    work page 2022

  18. [18]

    and Pfister, N., 2023

    Saengkyongam, S., Thams, N., Peters, J. and Pfister, N., 2023. Invariant policy learning: A causal perspective. IEEE transactions on pattern analysis and machine intelligence, 45(7), pp.8606-8620

    work page 2023

  19. [19]

    and Shpitser, I., 2023

    Richardson, T.S., Evans, R.J., Robins, J.M. and Shpitser, I., 2023. Nested Markov properties for acyclic directed mixed graphs. The Annals of Statistics, 51(1), pp.334-361

    work page 2023

  20. [20]

    Markov Properties for Graphical Models with Cycles and Latent Variables

    Forré, P. and Mooij, J.M., 2017. Markov properties for graphical models with cycles and latent variables. arXiv preprint arXiv:1710.08775

    work page Pith review arXiv 2017

  21. [21]

    Graphical aspects of causal models, Technical Report, R-191

    Verma, T., 1993. Graphical aspects of causal models, Technical Report, R-191. tech. rep., Cognitive Systems Laboratory, University of California at Los Angeles

    work page 1993

  22. [22]

    The Bayesian structural EM algorithm

    Friedman, N., 2013. The Bayesian structural EM algorithm. arXiv preprint arXiv:1301.7373. 40

    work page arXiv 2013